A NNALES DE L ’I. H. P., SECTION B
M. C RAMER
L. R ÜSCHENDORF
Convergence of a branching type recursion
Annales de l’I. H. P., section B, tome 32, n
o6 (1996), p. 725-741
<http://www.numdam.org/item?id=AIHPB_1996__32_6_725_0>
© Gauthier-Villars, 1996, tous droits réservés.
L’accès aux archives de la revue « Annales de l’I. H. P., section B » (http://www.elsevier.com/locate/anihpb) implique l’accord avec les condi- tions générales d’utilisation (http://www.numdam.org/conditions). Toute uti- lisation commerciale ou impression systématique est constitutive d’une infraction pénale. Toute copie ou impression de ce fichier doit conte- nir la présente mention de copyright.
Article numérisé dans le cadre du programme Numérisation de documents anciens mathématiques
http://www.numdam.org/
Convergence of
abranching type recursion
M. CRAMER and L.
RÜSCHENDORF
University of Freiburg, Institut fur Mathematische Stochastik,
Hebelstr. 27, Freiburg 79104, Germany.
Vol. 32, n° 6, 1996, p. 725-741 Probabilités et Statistiques
ABSTRACT. - The
asymptotic
distribution of abranching type
recursion isinvestigated.
The recursion isgiven by XiL(i)n-1
+Y,
where
( X L )
is a random sequence,( L ~L ~ 1 )
are iidcopies
ofLn-l,
K is arandom number and
.~, ( L ~ z ~ 1 ) , ~ ( X i ) , Y ~
areindependent.
This recursion has been studiedintensively
in the literature in the case that0,
K is nonrandom and Y = 0. Included in the moregeneral
recursion arebranching
processes, a model of Mandelbrot( 1974)
forstudying turbulence,
which was alsoinvestigated
in connection with infiniteparticle systems
and theexpansion
of total mass in the construction of random multifractalmeasures. The
stability
in the recursion arises from the fact that it includesa
smoothing part (addition)
and on the other hand apart increasing
thefluctuation
(random multiplier,
random number and randomimmigration).
Our treatment of this recursion is based on a contraction
technique
whichapplies
under some restrictions on the first two moments of the involved random variables. We obtain aquantitative approximation
result. Theexponential
rate of convergence can be observedempirically. Typically
after8 iterations the
limiting
distribution of the recursion is wellapproximated.
Key words: Branching type recursion, contraction method.
RESUME. - On etudie la distribution
asymptotique
d’une relation de recurrence detype
branchement donnée par + Y.Ici
(Xi)
est une suite de variables aléatoiresréelles,
sont desv.a.
indépendantes
de la meme distribution queL.,,-i,
K un nombreAMS Classification : 60 F 05, 68 C 25.
Annales de l’Institut Henri Poincaré - Probabilités et Statistiques - 0246-0203
726 M. CRAMER AND L. RUSCHENDORF
aléatoire et
K, (L~2~ 1 ) , ~ (XL ), Y~
sontindépendantes.
Dans la littératureon a etudie cette recurrence dans le cas où
Xi
>0,
K non-aléatoire et Y = 0. Entre autres la relation de recurrenceplus générale
décrit des processus debranchement,
un model de Mandelbrot(1974)
pour l’étude deturbulence,
et1’ expansion
de masse totale dans la construction desmesures aléatoires multifractales. La stabilité dans la recurrence vient du fait
qu’elle
contient unepartie
delissage (addition)
et conversement unepartie augmentant
la fluctuation(multiplicateur aléatoire,
nombrealéatoire et
immigration aléatoire). L’ analyse
dans cet article est fondéesur une
technique
de contractionapplicable
sousquelques
restrictions aux deuxpremiers
moments des v.a. données. Nous obtenons des résultatsd’ approximation quantitatifs.
Larapidité
de convergenceexponentielle
estobservable
empiriquement. Après
environ 8 iterations la distribution limite de la relation de recurrence est bienapprochée.
1. INTRODUCTION
Consider the
following
recursive sequence(Ln)
definedby
where are iid
copies
ofLn,_ 1,
is a real random sequence, K a random number inNo
and Y a randomimmigration
such thatK, ~ ( Xi ) . Y ~ . ( L~? ~ 1 )
areindependent. =
denotesequality
in distribution.(1)
induces a transformation T onM1,
the set ofprobability
distributionson
(R1, 31 ) , by letting T( )
be the distribution of03A3Ki=1 X ; 2i,
+V,
where(Zi)
are iid-distributed, (Zi), {(Xi), Y}, K independent.
Some
special
cases of this transformation resp. recursion have been studiedintensively
in the literature. If 1 then(1)
describes a Galton-Watson process with
immigration
Y and the number of descendants of an individuum describedby
K.(1)
can be considered from thispoint
of viewas a
branching
process with randommultiplicative weights.
Thespecial
casewhere K is constant, Y =
0. (Xi )
iid and nonnegative
has been introducedby
Mandelbrot(1974)
toanalyse
a model of turbulence ofYaglom
andKolmogorov.
This case has been studiedby
Kahane andPeyrière (1976)
Annales de l’Institut Henri Poincaré - Probabilités et Statistiques
and Guivarch
( 1990)
who considered thequestion
of nontrivial fixedpoints of T,
existence of moments of the fixedpoints
and convergence of(L~z ) .
ForXi -
the solutions of the fixedpoint equation Z a
are Paretian stable distributions
(if 0).
For this reason the solutionsare called semi-stable in Guivarch
( 1990).
In this paper we will bemainly
interested in the case of
multipliers Xi
and solutionsZi
with momentsof some order > 2. While the
analysis
of Kahane andPeyrière (1976)
isbased on an associated
martingale,
Guivarch( 1990)
uses a moreelementary martingale property together
with aconjugation
relation and momenttype
estimates for theLp-distance,
0 p 1.Motivated
by
someproblems
in infiniteparticle systems Holley
andLiggett (1981)
and Durrett andLiggett (1983)
considered this kind ofsmoothing
transformation with(Xi)
notnecessarily independent
and0,
K constant, Y = 0. In the last mentioned paper acomplete analysis
of this case could begiven.
Inparticular
a necessary and sufficient condition for the existence and the characterization of(all)
fixedpoints
and a
general
sufficient condition for convergence was derived as well as ageneralization
of the result of Kahane andPeyrière
on the existence ofmoments. The method of the paper of Durrett and
Liggett
is based on anassociated
branching
random walk.The use of contraction
properties
of minimalLp-metrics
in this paper allows to obtainquantitative approximation
results for the recursion(1).
Under the moment
assumptions
used in this paper the recursion convergesexponentially
fast to thelimiting
distribution. This is demonstratedby
simulations for several
examples.
Also it ispossible
to dismiss withthe
assumption
ofnonnegativity,
to deal with a random number and withimmigration
Y. This allows to includebranching
processapplications
as well as the consideration of the
development
of total mass in theconstruction of multifractal measures
(of
e.g. Arbeiter(1991)).
Thisexample
was
motivating
the work on this paper. Afterfinishing essentially
theinvestigation
on this paper we were informedby
G. Letac about thehistory
of the
problem.
We aregrateful
to him for his remarks and indications.In section 2 we consider the case Y -
0,
discuss a robustnessproperty
of theproblem,
relations toprevious
work andgive
some numericalexamples.
In section 3 we consider an extension to the case with
immigration
underassumptions
on theX,,
Y andLo assuring
thestationarity
of the first two moments. The method in this paper is based on a contraction method w.r.t.suitable metrics as
developed
in a sequence of furtherexamples
in Rachevand Ruschendorf
(1991).
728 M. CRAMER AND L. RUSCHENDORF
2. BRANCHING TYPE RECURSION WITH MULTIPLICATIVE WEIGHTS
In this section we consider the recursion
(1)
withpossibly dependent multipliers Xi
L andimmigration
Y -0,
i. e.where
L~2~ 1
are iidcopies
of(X L )
is a squareintegrable
realrandom sequence, K a random number in
No
andI~, ( X Z ) , (L~~)
areindependent.
To determine the correct normalization of
( L,~ )
we at first consider the first moments of(Ln). Denote in
:=ELn,
c := :=PROPOSITION 2.1. -
lo
=1, In
= C".Proof - Using
theindependence assumption
in(2)
and conditionalexpectations
we obtainSimilarly,
Annales de l’Institut Henri Poincaré - Probabilités et Statistiques
In the case b =
0,
vn = 0 for all n. Therefore we consideronly
thecase b > 0.
From
(3)
we obtain that for ac2,
is of the same order as This makes itpossible
to use thesimple
normalizationby In.
Define forc ~
0then
ELn =
1 andVar(Ln)
--~2
satisfies the modified recursion
where :=
L(i)n-1 cn-1.
Define
~2
to be the set of distributions on(tRB~)
with finite second moments and first momentequal
to one and define T :Ds
2014~~2 by
730 M. CRAMER AND L. RÜSCHENDORF
where
(Zi)
are i.i.d. random variables with distributionG, (X;, ) , ( 2; ) . K independent.
Letl2
denote the minimalL2-metric
onD2
definedby
where
F, G
are the distribution functions ofrespectively.
If ac2,
then T is a contraction w.r.t.
l ~ .
PROPOSITION 2.2. - Assume that a
c’,
thenfor F,
G ED2
Proof -
LetU ~ l > ‘~ F, ~,
~, E~,
be choosen on( S~ , A, P )
such thatIl~~l~ _ V~1>II~ _ ~~(F’, ~).d2
andK, (~~Tt), (Utl~, ~~1~), (U~2~. v~~2~), ...
independent.
ThenAs consequence of
Proposition
2.2 T hasexactly
one fixedpoint
inD2
with variance
equal
tob/ ( c~’
-a).
The fixedpoint equation
isgiven
inAnnales de l’Institut Henri Poincaré - Probabilites et Statistiques
terms of random variables
Z, Z,
ED2 , Z, (Zi) independent, by
and we obtain as
corollary:
In
particular Ln
converges in distribution to Z.PROPOSITION 2.4. -
If
K is constant andc~
V 2k h then x.
~
~~~
Proof - Ln
can beequivalently represented by Y,L
of thefollowing
formwhere
(X~l,....~h.-l,l ; ~~~, X~1,...,~~._l,h ) =(Xl, ~~~, Xh ) (c/: Guivarch, 1990).
(Yn )
is amartingale
and therefore is asubmartingale. Representing
theY,z
in the recursiveway Yn = ~ ~ ~ 1
whereY~~~ ~l
areindependent copies
ofY,,-i,
one obtainsOne can deduce from Theorem 2.3 that is
uniformly
bounded fork 2.
By
inductionover 1~ ~
one sees that the lower order terms in the aboveequation
areuniformly bounded,
sayby
C. Sinceone obtains
732 M. CRAMER AND L. RUSCHENDORF
Therefore,
theassumptions
of thisproposition
ensure that isuniformly
bounded for all k h. Thesubmartingale
convergence theoremnow
yields
the existence of anintegrable
almost sure limitof
SinceYn
the weak limit Z ofL.n
isabsolutely h-integrable.
DFor solutions of the
stationary equation (9)
it ispossible
to obtain astability
result in termsof lp
metrics defined as in(7)
with 2replaced by
p.
Suppose
we want toapproximate
the solution S of theequation
by
the solution of the"approximate" equation
where we assume
w.l.g.
c = 1 and consider the case ofindependent
sequences
( X i ) , ( X i ) ,
such that( X i ) , ( si )
and( X.L ) , ( Si )
areindependent
and K is constant.
PROPOSITION 2.5. -
If
constant,~,K 1 l p ( X i , X i )
~ and~~ 1 ~~Xi~~p l,
thenProof -
From the definition ofS,
S*Annales de l’lnstitut Henri Poincaré - Probabilités et Statistiques
This
implies
thatA similar idea for
establishing
robustness ofequations
can be found in Rachev(1991), Chapter
19.3.For random K we
replace Proposition
2.5by
Proof - By
thetriangle inequality
and theindependence assumption
andthe
assumption EXi
=EX,L
Therefore,
Remark. -
(a)
In the case of constant K andnonnegative Xi
Durrett andLiggett (1981) proved
that thestationary
solution Z of(9)
has momentsof order
j3
if andonly
ifFor
j3
=2, ( 14)
isequivalent
to the condition ac2
used in Theorem 1. In this sense this condition issharp
whenusing l2-distances.
Guivarch(1990)
has shown how to dismiss with the second moment
assumption.
734 M. CRAMER AND L. RUSCHENDORF
(b)
For the normalized recursion(5)
with(Xi ) i.i.d.,
K constant(where
we assume
w.l.g.
c =1),
we can use theexplicit
form(cf.
theproof
ofProposition 2.4)
where are
independent
and distributed asXi,
i.e.Ln
is thesum over
product weights
in thecomplete K-ary
tree. Fornonnegative multipliers Xi
one can consider further functionals as e.g.the max
product
over allpaths
oflength
n.Taking logarithms
and
application
ofKingman’s
subadditiveergodic
theoremyields
for someconstant
/3
This shows that in some sense the max
product weight
is notlarger
in theorder than the average
product weight.
For some cases,~~
isknown,
e.g. forX; a U ~0,1~ , ,~ ~
-0.23196(cf.
Mahmoud(1992),
p.165).
(c)
For some casesexplicit
solutions of(9)
are known.1 )
If K is constant,e X L d ~~ ( I , a - h )
isBeta-distributed,
thenZ ~ r(a, /~)
is Gamma-distributed(cf. Guivarch, 1990).
2)
IfXi 2L,
then with(Yi ) i.i.d., X d X1Z1
holdsConversely,
if~ h 1 Y; X 1,
then with(X.? ) i.i.d.,
the
( Zl )
solve(cf Durrett
andLiggett (1981)).
3)
If( ZL )
solve03A3Ki=1 X ; Zi , X,
i >0, then Yi
=Z1/03B8iWi, 0
03B82, WL
stable rv’s of index~,
solveAnnales de 1’Institut Henri Poincaré - Probabilités et Statistiques
To prove
(18), Z1/03B81W1 =
Yi.
Thisinteresting
transformationproperty
is used in Guivarch(1990)
to reduce the case of moments of
Xi
of low order to the case with moments ofhigher
order.4)
If~~ 1 X2 - c2 ~ 0,
thenZ ~ ~2) normally
distributed solve(9).
5)
If Z solves(9)
and Z is anindependent
copy ofZ,
then Z* := Z - Zsolves
whereXi
= andthe Ti
arearbitrary
randomsigns.
Inparticular
thecase K =
2, Xi a U~-1, l~ independent
is solvedby
Z* :- Z - Z whereZ a rC2~ 2 ~ o).
-(d)
Thefollowing
simulations(Figures
1 and2)
ofLn, K
=2, Xi,X2 independent, X 1 ~ X 2 d U ~0,1 ~ , respectively X i ~ ~3 ( 2, 2 )
showgood
coincidence with the theoretical Gamma-distribution.(e)
In the case K =2, Xl, X2 independent, X 1 a X? ~ U ~- g , g ~
noexplicit
solution of(9)
is known. Nevertheless thefollowing
simulations(Figure 3)
show thatLn
converges very fast to the fixedpoint
of(9).
Theempirical
distribution functions ofLio
andL12
canhardly
bedistinguished.
Therefore
they
may beregarded as
the limit distribution function. Theempirical
distribution function ofL6
isalready
very close to it.736 M. CRAMER AND L. RUSCHENDORF
(/) Branching
processes. -Equation (2)
includes the Galton-Watson process asspecial
case. A Galton Watson process is definedby
the recursionwhere X are i.i.d. with
reproduction
distribution inNo.
DefineK ~
X andXi - 1,
thenfor all n.
Annales de l’Institut Henri Poincaré - Probabilités et Statistiques
The
equality
can be seenby
induction in n. FirstZo
=Lo
= 1. IfZk == Lk
for k n, thenThe
assumption a cz
isequivalent
to the condition EX > 1. D Fromequality (20)
one can deriveexplicit stationary
distributions and extinctionprobabilities
in some cases. If e.g. X isgeometrically distributed,
P(X = k)
=p(l -
ENo,
thenand
Var(X)
= The normalized Galton-Watson processZ’ Zn Var
Zn)converges to a
(unique)
solution of the fixedpoint equation
The extinction
probability
iseasily
seen to be1 PP .
For the normalized continuouspart
anequation
identical to(21 ) (but
with differentvariances)
holds. It is well known that this
equation
is solvedby
thegeometric
stabledistribution of order
1,
i.e. theexponential
distribution. Thisimplies finally
since
3. A RANDOM IMMIGRATION TERM In this section we admit an additional
immigration
term.738 M. CRAMER AND L. RUSCHENDORF
Y}, K, L { 1 ~ 1, L ~ 2 ~ 2 ; ...
areindependent, X, and
Y have finitesecond moments. The
analysis
of(23)
isessentially simplified
if we assumefor
lo
:=ELo,
vo .-Var(Lo),
If c =
1,
then EY = 0 andlo arbitrary.
LEMMA 3.1. - Under
assumption (24)
holdsProof -
FromCondition
(24)
can be achieved with a twopoint
distribution forLo.
Itallows to use the
technique
ofproof
of section 2. Achange
of the initial condition leads to thenecessity
tochange
the method ofproof
and leadsto a
great variety
of different cases to be considered.We, therefore,
restrictto
(24)
in this paper.As in section 2 we introduce the
operator
Annales de l’Institut Henri Poincaré - Probabilités et Statistiques
where
M ( lo, vo)
is the set of distributions with meanlo
and variance r~and
(Vi)
arei.i.d., G, (VZ), {(X~).Y~.
Kindependent.
Similarly
as inProposition
2.2 we can show the contractionwhich
implies
the convergence ofLn
to theunique
fixedpoint
of T inM ( l o, vo )
w.r.t. thel2 -metric,
the contraction factorbeing ~.
We obtain a
sharper
result(i. e.
a smaller contractionfactor) by
the useof the
Zolotarev-metric (r
instead ofl2.
It is defined
by
PROPOSITION 3.2.
Proof -
Notethat (r
is ideal of order r w.r.t.summation,
i. e.for Z
independent
ofX,
Y andFor
(Z;);
i.i.d. distributedaccording
toF,G
we have with X =(Xz)
740 M. CRAMER AND L. RÜSCHENDORF
Note that
for
our recursion definedby
T the first two moments areconserved.
Therefore,
we canapply (29)
withr
3 and obtain ascorollary:
implies
where Z is a
fixed point of
T inM(lo; vo).
In
particular Ln
converges in distribution to Z.So also in the case with
immigration
one obtains anexponential
rateof convergence. As a consequence after a few iterations the
limiting
distribution is well
approximated.
Consider the
following example: L0 = 1 10 8-5 -f- 5 80 -+- 2 b2, K
=2,
Xi. X2 X d X 2 ~ U [- 2 , ] Y ~ ~-i
+~o + ~ ~2.
In this situation
(24)
is fulfilled. The fast convergence is confirmedby
the closeness of the
empirical
distribution functions ofL6
andL8
in thefollowing
simulation.FIG. 4. - Empirical distribution functions for Le and Ls .
Annales de l’Institut Henri Poincaré - Probabilités et Statistiques
REFERENCES
[1] M. A. ARBEITER, Random recursive constructions of self-similar fractal measures. The non-compact case. Prob. Th. Rel. Fields, Vol. 88, 1991, pp. 497-520.
[2] R. DURRETT and M. LIGGETT, Fixed points of the smoothing transformation.
Z. Wahrscheinlichkeitstheorie verw. Gebiete, Vol. 64, 1983, pp. 275-301.
[3] Y. GUIVARCH, Sur une extension de la notion de loi semi-stable. Ann. Inst. H. Poincaré, Vol. 26, 1990, pp. 261-286.
[4] R. HOLLEY and T.M. LIGGETT, Generalized potlach and smoothing processes.
Z. Wahrscheinlichkeitstheorie verw. Gebiete, Vol. 55, 1981, pp. 165-195.
[5] J. P. KAHANE and J. PEYRIÈRE, Sur certaines martingales de Benoit Mandelbrot. Adv. Math., Vol. 22, 1976, pp. 131-145.
[6] H. M. MAHMOUD, Evolution of Random Search Trees. Wiley, New York, 1992.
[7] B. MANDELBROT, Multiplications aléatoires itérées et distributions invariantes par moyenne
pondérée aléatoire. C. R. Acad. Sci. Paris, Vol. 278, 1974, pp. 289-292.
[8] S. T. RACHEV, Probability Metrics and the Stability of Stochastic Models. Wiley, New York, 1991.
[9] S. T. RACHEV and L. RÜSCHENDORF, Probability metrics and recursive algorithms. 1991, To
appear in Advances Appl. Prob., 1995.
(Manuscript received November 8, 1994;
revised September 1, 1995. )